Erfan Hardanian
All work

Sayasah-yah

Lets artificial intelligence work on a confidential file without seeing the names and numbers in it.

From sāye, Persian for shadow: it sees the shadow, never the thing.

PythonMCP serverPresidioClaude CodeOpen-sourced as obsifyOn GitHub

Field
Artificial intelligence and privacy
The shadow is set like a blank form: each box says what kind of thing goes in it, never what is there. Example data is made up.

Why I built it

At work, people held back from artificial intelligence over privacy. Yet without it, it’s hard to compete.

I wanted to give them a safe way in, so I built Saya and shared it with them. To write the code, a model needs only the shape of a file, never its values.

The regulator’s advice

As a matter of best practice, the OAIC recommends that organisations do not enter personal information, and particularly sensitive information, into publicly available generative AI tools, due to the significant and complex privacy risks involved.

OAIC, 2024

How it works

ModelCode

Saya never calls a model itself. It stops accidents, not a determined attacker, and its README says so.

  1. Guard

    A file marked sensitive can’t be opened directly; the assistant is sent to Saya instead.

    Code

  2. Scan

    Finds names, tax file numbers, company numbers and passwords, and reports only what kind and how many. In an outside test of 300 records it caught 87%; a miss can only slip out through the returned output, which is why only totals come back.

    Code

  3. Twin

    A stand-in file with the same columns, types and row count, every value made up.

    Code

  4. Write

    The model writes its code against the stand-in, where a mistake costs nothing.

    Model

  5. Run

    The code runs on the real file, on your machine, network off.

    Code

  6. Return

    Only the output comes back, masked and cut to size.

    Code

Built with

Runs on
Python, as an MCP server on your own computer, with no network access. Reads spreadsheets, PDFs and Word files.
Size
4,400 lines of Python. Free and open source (MIT), on PyPI and Docker.
Tooling
Claude Code.

The rule it follows

Share the shape. Keep the values where they are.